Add `riscv info` command. Final output is "TCL format" and looks like this:
```
hart.xlen 64
hart.trigger_count 4
dm.abits 6
dm.progbufsize 2
dm.sbversion 0
dm.sbasize 0
dm.sbaccess128 0
dm.sbaccess64 0
dm.sbaccess32 0
dm.sbaccess16 0
dm.sbaccess8 0
```
* Add `riscv info` command.
This command displays some basic information that OpenOCD has detected
about the target. The output is displayed in YAML so it can easily be
parsed. Example of current output:
```
Hart:
XLEN: 32
trigger count: 4
Debug Module:
abits: 6
progbufsize: 2
sbversion: 0
sbasize: 0
sbaccess128: 0
sbaccess64: 0
sbaccess32: 0
sbaccess16: 0
sbaccess8: 0
```
Change-Id: If920c083ff6ec9f482c50f913cd8ceaa62461217
Signed-off-by: Tim Newsome <tim@sifive.com>
* Disable workflow inherited from upstream.
Change-Id: Ifc5ed1b4f5ec2278b8bcf3279c9fd462e469fefa
Signed-off-by: Tim Newsome <tim@sifive.com>
* Switch from YAML to TCL "set array" input format.
Change-Id: I3833210e5bf6d7cffc9934c04ec5201ae7732ad8
Signed-off-by: Tim Newsome <tim@sifive.com>
* Remove indent in `riscv info` output.
That was getting a little too cute, and probably more confusing than
helpful.
Change-Id: Ie51416f53ab4b69294962f0565767d370db82867
Signed-off-by: Tim Newsome <tim@sifive.com>
* Fix error handling in read_memory_progbuf_one().
Be sure to restore mstatus/s0 even if there is a failure during the
operation.
Fixes#559.
Change-Id: Ib86ca2c7455bad4a668f34703566060a782116db
Signed-off-by: Tim Newsome <tim@sifive.com>
* Style fix suggested in review.
Change-Id: I444112a9dffea483b7d0e5f96ef7bbdaf58d249f
Signed-off-by: Tim Newsome <tim@sifive.com>
This is a fix for an issue reported by Joe Stoy at:
https://sourceforge.net/p/openocd/mailman/message/37128537/
When clearing sbcs.sbbusyerror, preserve other bits in the sbcs
register that are needed for subsequent system bus transactions.
This lets a user see exactly what period of time was sampled, without
having to guess how much time the target was ignored in between bursts.
Change-Id: I5c0639528636bf1a88f249be3ba59bec28c001e2
Signed-off-by: Tim Newsome <tim@sifive.com>
Used histogram diff strategy, which was much better than the default.
Conflicts:
doc/openocd.texi
src/flash/nor/fespi.c
src/jtag/drivers/libjaylink
src/rtos/rtos.c
src/target/riscv/batch.c
src/target/riscv/encoding.h
src/target/riscv/riscv-011.c
src/target/riscv/riscv-013.c
src/target/riscv/riscv.c
src/target/riscv/riscv.h
src/target/target.c
tcl/target/gd32vf103.cfg
Change-Id: I1321f62ba719419e58f93b2195f2540bd62f50d2
These are all the changes from https://github.com/riscv/riscv-openocd
(approximately 91dc0c0c) made just to src/target/riscv/*. Some of the
new code is disabled because it requires some other target-independent
changes which I didn't want to include here.
Built like this, OpenOCD passes:
* All single-RV32 tests against spike.
* All single-RV64 tests against spike.
* Enough HiFive1 tests. (I suspect the failures are due to the test
suite rotting.)
* Many dual-RV32 (-rtos hwthread) against spike.
* Many dual-RV64 (-rtos hwthread) against spike.
I suspect this is an overall improvement compared to what's in mainline
right now, and it gets me a lot closer to getting all the riscv-openocd
work upstreamed.
Change-Id: Ide2f80c9397400780ff6780d78a206bc6a6e2f98
Signed-off-by: Tim Newsome <tim@sifive.com>
Reviewed-on: http://openocd.zylin.com/5821
Tested-by: jenkins
Reviewed-by: Jan Matyas <matyas@codasip.com>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-by: Karl Palsson <karlp@tweak.net.au>
* Add memory sampling feature.
Currently only gets 10 samples per second, but the overall scaffolding
looks like it works.
Change-Id: I25a2bbcba322f2101c3de598c225f83c902680fa
* Basic memory sample speed-ups.
977 samples/second.
Change-Id: I6ea874f25051aca1cbe3aa2918567a4ee316c4be
* Add base64 dumping of sample buffer.
We can't just dump raw data, because the API we use to get data to the
"user" uses NULL-terminated strings.
Change-Id: I3f33faaa485a74735c13cdaad685e336c1e2095f
* WIP on optimizing PC sampling.
1k samples per second on my laptop, which is roughly double what it was.
Change-Id: I6a77df8aa53118e44928f96d22210df84be45eda
* WIP
Change-Id: I4300692355cb0cf997ec59ab5ca71543b295abb0
* Use small batch to sample memory.
5k samples/second. No error checking.
Change-Id: I8a7f08e49cb153699021e27f8006beb0e6db70ee
* Collect memory samples near continuously.
Rewrite OpenOCD's core loop to get rid of the fixed 100ms delay.
Now collecting 15k samples/second.
Change-Id: Iba5e73e96e8d226a0b5777ecac19453c152dc634
* Fix build.
Change-Id: If2fe7a0c77e0d6545c93fa0d4a013c50a9b9d896
* Fix the mess I left after resolving conflicts.
Change-Id: I96abd47a7834bf8f5e005ba63020f0a0cc429548
* Support 64-bit address in memory sampling.
* Support sampling 64-bit values.
* Better error reporting. WIP on 64-bit support.
* Speed up single 32-bit memory sample.
21k samples/second.
* WIP on review feedback.
Change-Id: I00e453fd685d173b0206d925090beb06c1f057ca
* Make memory sample buffers/config per-target.
Change-Id: I5c2f5997795c7a434e71b36ca4c712623daf993c
* Document, and add bucket clear option.
Change-Id: I922b883adfa787fb4f5a894db872d04fda126cbd
Signed-off-by: Tim Newsome <tim@sifive.com>
* Fix whitespace.
Change-Id: Iabfeb0068d7138d9b252ac127d1b1f949cf19632
Signed-off-by: Tim Newsome <tim@sifive.com>
* Document sample buffer full behavior.
Change-Id: Ib3c30d34b1f9f30cf403afda8fdccb850bc8b4df
Signed-off-by: Tim Newsome <tim@sifive.com>
* Actually clear the sample buffer in dump_sample_buf.
Change-Id: Ifda22643f1e58f69a6382abc90474659d7330ac5
Signed-off-by: Tim Newsome <tim@sifive.com>
* Use compatible string formatting.
Change-Id: Ia5e5333e036c1dbe457bc977fcee41983b9a9b77
Signed-off-by: Tim Newsome <tim@sifive.com>
* riscv: work around buggy hart states during reset in some DMs
As described in the comment this change adds, the GD32VF103 DM reports
that the hart is in more than one state while it is resetting. Because
of this, the current code acknowledges resets before they actually
complete. This sometimes prevents havereset from getting cleared as
intended, leading to a spurious "Hart 0 unexpectedly reset!" message the
next time riscv_is_halted() gets called.
To work around this, check for the absence of the unavailable state
rather than the presence of the running or halted states. This behavior
is also arguably more true to the spec than what exists now: Section 3.2
states that "The system may take an arbitrarily long time to come out of
reset, as reported by allunavail, anyunavail."
Change-Id: I34e90a16233125608bce8e4c2414dbead637600e
Signed-off-by: Thomas Hebb <tommyhebb@gmail.com>
* riscv: support custom reset-assert scripts
The reset-assert event is used, if present, to override the default
reset logic for ARM and MIPS cores. Do the same for RISC-V so that
devices with buggy ndmreset functionality (like GD32VF103) or
nonstandard reset sequences can specify the appropriate logic in Tcl.
Change-Id: I5e12077d67509853edb8ef3ad3f037f293a5fbb6
Signed-off-by: Thomas Hebb <tommyhebb@gmail.com>
* tcl/target: support GD32VF103 RISC-V MCU
The GD32VF103 is a low-cost 32-bit RISC-V microcontroller with
peripherals that are more-or-less compatible with the STM32F103 ARM
microcontroller. It is available on several low-cost dev boards, such as
the Sipeed Longan Nano, which is what I am testing on.
Add initial support for this chip, including a workaround for a buggy
ndmreset line (i.e. one that doesn't actually trigger a reset) in its
integrated debug module. Use the existing GD32VF103 flash driver that
was ported from the vendor's code in commit 48e40f3513 ("Add support
for GD32VF103 flash").
Change-Id: Iadac47ceb5437b8e18f3d35901388f10fef9f876
Signed-off-by: Thomas Hebb <tommyhebb@gmail.com>
* tcl/target/gd32vf103: add main flash alias
The GD32VF103 creates an alias to either main flash or the bootloader at
0x0, depending on how it was booted. As such, we want to indicate to
debuggers that the memory at 0x0 is flash and so cannot support software
breakpoints. To do this, add an alias to the main flash in the config.
This isn't strictly accurate in the case where we're running the
bootloader, but it still suits our purpose of fixing breakpoint
behavior.
Change-Id: I9eb8462d354f096eee231c0e5e2bffa538a5903e
Signed-off-by: Thomas Hebb <tommyhebb@gmail.com>
This seems to be completely unused in these two files. It was probably
accidentally copied from riscv-011.c, where it is used.
Change-Id: I3f7ad8b2d26b005d3ea4438e2b3ec46a6c801792
Signed-off-by: Thomas Hebb <tommyhebb@gmail.com>
* Allocate RISC-V arch_info during target creation
* Ensured that target->arch_info is allocated as soon as the
target is created. Needed so that per-target config commands
(e.g. "riscv set_mem_access") can be executed also in the
OpenOCD's config phase (before calling "init").
* Added several assert()'s for safety.
Signed-off-by: Jan Matyas <matyas@codasip.com>
* Removed a TODO comment
* Add flexible selection of memory access methods, detection of aampostincrement.
New configuration command introduced: "riscv set_mem_access".
It allows to specify which RISC-V memory access methods (progbuf,
sysbus and/or abstract access) should be tried and in which order
of priority.
Command "riscv set_prefer_sba" is left and works in backward
compatible way, but is marked as deprecated.
First time abstract memory access is executed, it is tried with
set aampostincrement bit. If the abstract command fails or the
address is not incremented correctly, aampostincrement will not
be used for any subsequent accesses.
Signed-off-by: Samuel Obuch <sobuch@codasip.com>
* remove unnecessary variable
* fix doc
A ton of constants got a new prefix, so I made a lot of changes to
match, but no functional changes.
I did define DTM_DMI_MAX_ADDRESS_LENGTH in batch.c. That definition
never should have been in debug_defines.h, which I missed during code
review.
Change-Id: If5d86660f84bb0a3f2865fb36ef05d6630486d8b
Accessing registers on targets that implement 0.9 or earlier will no
longer work. If you need that we can talk about making it a config
option.
Change-Id: I953b639cf9a92ee9b0422e035da57c1d07504237
* WIP, apply stash with conflicts.
Change-Id: Ia794bde419aa29161c68898d20e30527e69f5a31
* Fix conflict resolution problems.
Change-Id: I4cedc348cf613f98cc5a36886f37c568ca644238
* Add repeat_read command.
Only implemented for sba v1 right now, and poorly tested at that.
Change-Id: I1d9ff63e1dea14b3f6a9f8ba4dad53668bf8038b
* Hide bogus address in repeat_read
Change-Id: Ib66c1fa60df9c7fc7cc87880b0fddc52825b48aa
* WIP make repeat read work with progbuf.
Change-Id: I555f8b880c8bf0d1ed0f3f90c7987a5b516a7a79
* WIP
Change-Id: Ic567cea68355ae907e94bd25185a2c9be6fd798d
* Fix error handling when increment is non-zero.
Change-Id: I5a2f3f2ee948fd4e12c0443a542e85b7b5c5791a
* Correctly(?) handle failures when increment is 0.
I'm not 100% convinced that this ensures every read value shows up in
the output, but it ought to work.
Change-Id: I1af3e7174cf9d5e6f293456fb5ead629e17faaaa
* Don't crash when asked to read no data.
Change-Id: I4061b5c720a43a4f828384ab9eacc89557adfa05
* Remove unnecessary comment.
Change-Id: I1be3d699b86299339b3a830ca1ef13c9f5b9fe0f
* Document `riscv repeat_read`.
Change-Id: I4a0f071f38784b2de034f8c1b0ce75d6d2d326b2
* WIP making semihosting work with -rtos hwthread.
Change-Id: Icb46f3eeedc1391e8fdc73c3ad8036f20267eb2e
* More WIP.
Change-Id: I670a6e1ba2a13a6ef2ae303a99559a16fdd1bbfb
* Fix halting due to a trigger.
Change-Id: Ie7caa8dde9518bcd5440e34cf31ed0d30ebf29ad
* Fix multicore semihosting without halt groups.
Change-Id: I53587e5234308ed2cc30a7132c86e4c94eb176c4
* WIP
Change-Id: I40630543b08d8b533726cb3f63aa60a62be8ef40
* Fix single core semihosting.
This was the last bug!
Change-Id: I593abac027fa9707f48b7f58163d7089574a0e28
* Fix whitespace.
Change-Id: I285c152970b87864c63803fae61312e5b79dfe6d
Line continuation, adding a backslash as last char of the line, is
requested in multi-line macro definition, but is not necessary in
the rest of C code.
Remove it where present.
Identified by checkpatch script from Linux kernel v5.1 using the
command
find src/ -type f -exec ./tools/scripts/checkpatch.pl \
-q --types LINE_CONTINUATIONS -f {} \;
Change-Id: Id0c69e93456731717a7b290b16580e9f8ae741bc
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: http://openocd.zylin.com/5619
Reviewed-by: Andreas Fritiofson <andreas.fritiofson@gmail.com>
Tested-by: jenkins
In a switch/case statement, a break placed after a goto or return
is never executed.
The script checkpatch available in Linux kernel v5.1 issues a
warning for such unused break statements.
In the process of reviewing the new checkpatch for its inclusion
in OpenOCD, let's get rid of these warnings.
The script checkpatch is unable to fixup automatically this case.
Thanks to having "break" command using a single code line, this
patch has been generated using the script below:
find src/ -type f -exec ./tools/scripts/checkpatch.pl -q \
--types UNNECESSARY_BREAK -f {} \; \
| sed -n '/^#/{s/^.*FILE: //;s/:$//;s/:/ /;p}' \
| awk 'function P() {print "sed -i '\''"b"'\'' "a};
{
if ($1!=a) {
if (a) {P()};
a=$1;
b=$2"{d}";
} else {
b=b";"$2"{d}"
}
};
END {P()}'
Change-Id: I56ca098faa5fe8d1e3f712dc0a029a3f10559d99
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: http://openocd.zylin.com/5617
Reviewed-by: Andreas Fritiofson <andreas.fritiofson@gmail.com>
Tested-by: jenkins
BUILD_TARGET64 creates a larger test matrix and mostly gates the
building of the aarch64/armv8 target, make that unconditional, which
would help fixing any issues with 64-bit address types anyway.
Rebased by Antonio Borneo after commit 1fbe8450a9 ("mips: Add
MIPS64 support")
Change-Id: I219f62b744d540d9dde9a42e6b63fd7d91df3dbb
Suggested-by: Matthias Welwarsky <matthias@welwarsky.de>
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: http://openocd.zylin.com/5240
Tested-by: jenkins
Instead of exiting during examine(), spit out a warning, and don't
expose the vector data registers. We do provide access to the vector
CSRs, because maybe they do work? It's just that we have no idea what
the size of the data registers is.
Change-Id: I6e9ffeb242e2e22fc62cb1b50782c2efb4ace0bd
Writing bits to an uninitialized buffer generated false warnings.
Zero buffers before setting them by buf_set_u32|64()
(do it only if bit-by-bit copy loop is used,
zeroed buffer is not necessary if a fast path write is used)
Change-Id: I2f7f8ddb45b0cbd08d3e249534fc51f4b5cc6694
Signed-off-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-on: http://openocd.zylin.com/5383
Tested-by: jenkins
Reviewed-by: Andreas Fritiofson <andreas.fritiofson@gmail.com>
The script checkpatch available in new Linux kernel offers an
experimental feature for automatically fix the code in place.
While still experimental, the feature works quite well for simple
fixes, like spacing.
This patch has been created automatically with the script under
review for inclusion in OpenOCD, using the command
find src/ -type f -exec ./tools/scripts/checkpatch.pl \
-q --types POINTER_LOCATION --fix-inplace -f {} \;
then manually reviewed.
OpenOCD coding style does not mention the space around pointer's
asterisk, so no check is enforced. This patch only makes the style
uniform across the files.
The patch only changes amount and position of whitespace, thus
the following commands show empty diff
git diff -w
git log -w -p
git log -w --stat
Change-Id: Iefb4998e69bebdfe0d1ae65cadfc8d2c4f166d13
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: http://openocd.zylin.com/5197
Tested-by: jenkins
* WIP
Change-Id: I0264a73b7f7d2ce89cc0b80692dbf81d9cdcc2fd
* Reading v* registers appears to work.
Can't really test it though, because gdb doesn't print them right.
Change-Id: I8d66339371c564a493d32f15c3d114b738a455c5
* Total hack to communicate registers to gdb.
Change-Id: Id06c819675f2a5bcaf751e322d95a7d71c633765
* Implement writing vector registers.
Fixed reading vector registers.
Change-Id: I8f06aa5ee5020b3213a4f68644c205c9d6b9d214
* Show gdb the actual size of the vector registers.
This length may be different per hart.
Change-Id: I92e95383da82ee7a5c995822a53d51b1ea933493
* Remove outdated todo comment.
Change-Id: Ic9158b002858f0d15a6452773b095aa5f4501128
* Removed TODO comment.
Filed #449 to track this.
Change-Id: I5277b19e545df2024f34cda39158ddf7d0d89d47
* Nicely handle some errors reading/writing V regs.
Change-Id: Ia7bb63a5f9433d9f7b46496b2c0994864cfc4a09
* Handle DMI busy in sba write.
If we encounter DMI busy on the NOP after a read, we'll never get the
value out because DMI busy is sticky. The read must be retried, but we
don't know whether it was ever issued. Since the read has side effects
(incrementing of the address) this retry must be handled at a higher
layer. So now dmi_op_timeout can be told to retry or not, and if retry
is disabled it'll return an error when busy.
Also actually properly do the retry in dmi_op_timeout(). Previously the
code would not reissue the command and end up returning a garbage value.
Change-Id: I3b52ebd51ebbbedd6e425676ac861b57fbe711b1
* Fix whitespace.
Change-Id: Icb76d964e681b22346368d224d1930c9342343f3
* Handle a few more DMI busy cases.
Change-Id: I8503a44e4bf935c0ebfff0d598fe4c322fda702a
* Explain when to use dmi_op_timeout(retry).
Change-Id: I1a5c6d76ac41a84472a8f79faecb2f48105191ff
* dmi_reset does not affect the current transaction.
That means the retry scheme we had been using works fine. This does
contain some minor tweaks, and now we pass my tests which hammer the DMI
busy case harder.
Change-Id: I13eee384dbba82bc5a5b1d387c75c547afe557b5
* Remove unnecessary changes to make the PR readable
Change-Id: I87079876e6965563cf590e3936b3595aeab8715d
* Move idle to end of line...
... because we go through run-test/idle after the scan.
Change-Id: I21a8cff22471f0b895d8cd8d25373dced9bf1ca9
* Remove unused code.
Change-Id: I07a7cdd2d64ca40a4fe181111a34cf55ff1928d1
The original OpenOCD code issued FENCE & FENCE.i twice for the current
hart (which is harmless, but takes time).
Avoiding this extra FENCE is a slight performance improvement. Per my rough
measurements, this improves performance of certain debugger actions
(single-stepping) by approx. 20% in single-hart systems.
Abstract write size (aarsize) to shall always match the real
size of the register. This is because abstract write of smaller size
than the register need not be supported per spec (pg. 13 of RISC-V
External Debug Support ver. 0.13.2).
Fix memory access on 64-bit targets with no progbuf and sba that
supports 32-bit accesses but not 64-bit accesses. Bug was introduced in #419.
This fixes https://github.com/riscv/riscv-tests/issues/217.
Change-Id: Ib5ddf9886b77e3d58fe1d891b560ad03d5a46da1
When allocating scratch memory within RISC-V target
(scratch_reserve()), take into account whether progbuf
is writable or not, as determined by examine_progbuf().
* fix for batch scans not honoring presence of BSCAN tunnel
* fix formatting to placate checkpatch
* replace DIM with ARRAY_SIZE
* Refactor code that adds a bscan tunneled scan.
* Move bscan tunnel context to the batch structure, and in array
form, one per scan
* adjust code that was inconsistent with project code formatting standards
* 64-bit progbuf memory reads work.
Change-Id: Ia3dbc0ee39a31ed0e5c38bbb3d9e089b2533f399
* 64-bit writes work.
Change-Id: Iae78711d715b6682817bb7cce366b0094bda8b23
* Let targets indicate number of supported data bits.
This is used by the default memory read/write functions when creating an
aligned block.
I'm adding this mainly to ensure I get coverage of the 64-bit progbuf
memory read/write code.
Change-Id: Ie5909fe537c9ec3360a8d2837f84be00a63de77b
* Make mingw32 happy.
Change-Id: Iade8c1fdfc72ccafc82f2f34923577032b668916
* Add riscv_batch_available_scans().
This function will query the number of available scans in a batch.
* Perform SBA writes with batch transactions for improved performance.
Using batch transactions avoids an unnecessary dmi read after every
dmi write, resulting in a significant performance improvement.
Tim Newsome
Jan Matyas
Tom Hebb
Samuel Obuch
Samuel Obuch
Antonio Borneo
Florian Fainelli
Tomas Vanek
Greg Savin
darius-bluespec